半监督系数选择法的人脸识别

A semi-supervised coefficient selection method for face recognition

针对人脸识别过程中图像数据维数过高以及需要大量类别标记的问题,提出一种半监督离散余弦变换系数选择法,用以实现数据降维并提高识别率.该算法首先将图像数据进行离散余弦变换,根据频率特征通过预掩模选取有用信息;然后进行半监督约束聚类,利用少量有标记样本的约束集,对训练图像进行聚类;根据类别搜索较高的判别系数值,获得系数选择掩模以及训练图像的投影阵.将测试图像离散余弦变换阵在此掩模上投影,计算其与训练图像投影阵距离,利用分类器进行分类.在ORL与Yale人脸数据库上的实验结果表明:所提方法的性能优于传统方法,并与主成分分析与线性判别分析进行组合,获得了90%以上的识别率.

In face recognition, there exist the problems of high dimensionality of image data and requiring many class labels. To overcome these, a semi-supervised discrete cosine transform （DCT） coefficient selection method was proposed to reduce dimensions and improve recognition accuracy. First, the DCT was performed on an image database, and useful features were selected by pre-masking based on frequency features. Second, with a few labeled samples, semi-supervised constrained clustering was used for clustering on training image sets. Then, higher discriminant coefficient values were obtained by class labels, and coefficient selection masking and projection of training images were carried out. Finally, the discrete cosine transform the test images was projected on the masking, the distance between test image projection and training image projection was computed, and a class of test images was estimated and classified based on the minimal distance classifier. Experimental results on ORL and Yale face databases show that the performance of the proposed method is better than traditional methods. Furthermore, the proposed method can be combined with principal component analysis （PCA） or linear diseriminant analysis （LDA）, and obtain more than 90% recognition rate.